ComfyUI in 2026: Set Up a Local Product-Image Studio on a 12 GB GPU
By the end of this post you will have ComfyUI running on your own machine with a commercially licensed image model, a template workflow for product shots, and a one-screen interface your colleagues can use without seeing a node graph. The cheapest hardware that does this comfortably is a used 12 GB GPU or a 16 GB Apple Silicon Mac; a laptop with no GPU can follow along in CPU mode, slowly.
What you need
- Hardware: an NVIDIA GPU with 8 GB or more (12 GB is comfortable), an Apple Silicon Mac with 16 GB, or an AMD/Intel GPU with the caveats below. The official system requirements list NVIDIA (PyTorch with CUDA 13.0), AMD (ROCm 7.2), Intel Arc (PyTorch
torch.xpu), Apple Silicon (Metal) and a--cpufallback. - Disk: the desktop app needs “at least 4.85 GB of disk space and 8 GB RAM (16 GB recommended)” per the desktop install guide; add 10 to 15 GB for models.
- Software: ComfyUI Desktop (Windows 10+, macOS 13+ on Apple Silicon) or the manual install on Linux. Python 3.13 is recommended, 3.12 works.
- Time: 30 to 45 minutes including downloads. No account, no API key.
What the ecosystem looks like in 2026
ComfyUI stopped being a single Python script a while ago. The pieces you will meet:
| Piece | What it is | Where we stand on 2026-09-09 |
|---|---|---|
| Core | The Python server that runs workflows | Upstream v0.34.0 released 26 August 2026; our workstation runs 0.25.0 |
| Desktop app | Installer that bundles Python, the server and updates | v0.9.4, 28 May 2026; Windows and macOS only, Linux from source |
| Frontend | The graph editor, shipped as a package | 1.45.19 on our machine, per /system_stats |
| Manager | Installs and updates custom node packs from the UI | Ships in custom_nodes/ComfyUI-Manager on our install |
| Templates | Ready-made workflows for each model family | comfyui-workflow-templates 0.10.0 here |
| App Mode, App Builder, ComfyHub | Hide the graph, expose three inputs, share as a URL | Announced 10 March 2026 (ComfyUI blog, press release) |
Node count is a good measure of how much the platform has grown. Our install answers /object_info with 1,035 node types: 559 core, 232 “partner” nodes that call paid cloud APIs (OpenAI, Kling, Runway and others) and 244 from custom packs, mostly KJNodes. You never have to touch the partner nodes; everything in this post runs offline.
Model formats are simpler than they look. Weights come as .safetensors; a suffix such as fp8 means the file was stored at 8-bit precision to halve memory, and community GGUF builds exist for very tight VRAM. Files go into fixed folders: models/checkpoints/ for all-in-one SD-class files, models/diffusion_models/, models/text_encoders/ and models/vae/ for newer models shipped in parts.
Step 1: Install
Windows or macOS: download the desktop app from comfy.org, run it, let it create its Python environment. Linux (what we run):
git clone https://github.com/Comfy-Org/ComfyUI ~/dev/ComfyUI && cd ~/dev/ComfyUI
python3 -m venv .venv && source .venv/bin/activate
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130 # NVIDIA
pip install -r requirements.txt
python main.py --listen 127.0.0.1 --port 8188Open http://127.0.0.1:8188. Check what the server sees:
curl -s localhost:8188/system_stats | python3 -m json.tool | head -30On our workstation this reports comfyui_version 0.25.0, pytorch_version 2.12.1+cu130, Python 3.12.3 and one CUDA device, an NVIDIA GB10 with 130 GB visible memory. If devices shows only cpu, PyTorch was installed without GPU support: reinstall it with the index URL for your platform from the requirements page. No GPU at all? Add --cpu; images take minutes instead of seconds, which is fine for testing a workflow before you buy hardware.
Step 2: Pick a model you may use commercially
For product images the choice in 2026 is FLUX.2 [klein] 4B. It is the one FLUX.2 weight under the Apache 2.0 licence, so commercial use is allowed; the 9B and [dev] variants are non-commercial (licence summary). The official ComfyUI tutorial lists the files and the VRAM measured by the ComfyUI team:
| File | Folder | Notes |
|---|---|---|
flux-2-klein-4b-fp8.safetensors (distilled) | models/diffusion_models/ | ”~1.2s (5090) · 8.4GB VRAM” |
flux-2-klein-base-4b-fp8.safetensors (base) | models/diffusion_models/ | ”~17s (5090) · 9.2GB VRAM” |
qwen_3_4b.safetensors | models/text_encoders/ | text encoder |
flux2-vae.safetensors | models/vae/ | decoder |
8.4 GB of VRAM is why a 12 GB card is the comfortable floor and an 8 GB card is possible with offloading. If you only have 8 GB and want speed, Stable Diffusion 1.5 still works: our own install carries v1-5-pruned-emaonly-fp16.safetensors (4.27 GB) in models/checkpoints/, and 641 images have come out of that output/ folder so far.
Step 3: Load the template and generate
In the interface open Workflow > Browse Templates, search “Flux.2 Klein”, and pick the 4B distilled template. Missing files are flagged in red with a download link; put them in the folders above and press R to refresh. Type a prompt such as “white ceramic mug on a light oak table, soft window light, product photo, 1:1”, queue it, and check the image lands in output/. Expected time on a 12 GB card: seconds per image, not minutes.
Step 4: Hide the graph for your colleagues
Click Enter App Mode in the top bar. In App Builder, expose only the prompt, the seed and the image size; everything else keeps the values you set. Save the workflow: anyone opening it sees a form with three fields and a result panel. This is the single feature that turns ComfyUI from a designer’s tool into something a shop assistant can use on a Monday morning.
Step 5: Automate from a script
The server is an HTTP API. Export the workflow with Export (API) and queue it from Python:
import json, requests
wf = json.load(open("product-shot-api.json"))
wf["6"]["inputs"]["text"] = "leather wallet, top view, white background, product photo"
r = requests.post("http://127.0.0.1:8188/prompt", json={"prompt": wf})
print(r.json()["prompt_id"]) # results appear under /history/<prompt_id>Point a loop at your product catalogue and you have the batch pipeline from our batch generation tutorial. For consistent angles across a range, add the edge-guided step from our ControlNet tutorial.
What we measured and what we could not
| Item | Value on our workstation | Source |
|---|---|---|
| ComfyUI core / frontend / templates | 0.25.0 / 1.45.19 / 0.10.0 | /system_stats, 2026-09-09 |
| PyTorch / Python | 2.12.1+cu130 / 3.12.3 | /system_stats |
| Node types | 1,035 (559 core, 232 partner, 244 custom) | /object_info |
| Custom packs installed | ComfyUI-Manager, KJNodes, WASasquatch | ls custom_nodes/ |
| Klein 4B VRAM | 8.4 GB distilled, 9.2 GB base | ComfyUI docs, measured on an RTX 5090 |
We have not yet timed FLUX.2 [klein] on a 12 GB card ourselves; the seconds-per-image figure above is the ComfyUI team’s on a 5090, and a 12 GB card will be slower. We will publish our own timing when we run it on client hardware.
Where this fits
A local image studio replaces per-image cloud fees and keeps unreleased products off third-party servers. Honest limits: a 4B model will not match the largest hosted models on complex scenes with text, hands or several people; for a catalogue of single objects on clean backgrounds it is more than enough, and the licence lets you ship the result.
Next steps
- Generate a hundred variants from a CSV: ComfyUI batch image generation.
- Keep the same angle across a product range: ComfyUI ControlNet tutorial.
- Choose the GPU or Mac before you spend: Edge AI hardware guide 2026.
Work with us
We run ComfyUI in production on our own hardware and set it up for clients with the licence question answered before the first image. If you want a product-image pipeline that stays in your building, get in touch or see how our consulting works.